Hitotsubashi University Repository
Title
Spillovers as a Driver to Reduce Ex-post Inequality
Generated by Randomized Experiments: Experiments
from an Agricultural Training Intervention
Author(s) TAKAHASHI, Kazushi; MANO, Yukichi; OTSUKA, Keijiro
Citation
Issue Date 2018-05
Type Technical Report
Text Version publisher
URL http://hdl.handle.net/10086/29259
Right
HIAS-E-69
Spillovers as a Driver to Reduce Ex-post Inequality Generatedby Randomized Experiments:
Experiments from an Agricultural Training Intervention
Kazushi TakahashiSophia University, Faculty of Economics, Tokyo 102-8554, Japan
Yukichi ManoHitotsubashi University, Graduate School of Economics, Tokyo 186-8601, Japan
Keijiro OtsukaKobe University, Graduate School of Economics, Hyogo 657-8501, Japan
May 2018
Hitotsubashi Institute for Advanced Study, Hitotsubashi University2-1, Naka, Kunitachi, Tokyo 186-8601, Japan
tel:+81 42 580 8604 http://hias.ad.hit-u.ac.jp/
HIAS discussion papers can be downloaded without charge from:http://hdl.handle.net/10086/27202
https://ideas.repec.org/s/hit/hiasdp.html
All rights reserved.
1
Spillovers as a Driver to Reduce Ex-post Inequality Generated by Randomized
Experiments: Evidence from an Agricultural Training Intervention*
Kazushi Takahashi, Yukichi Mano, and Keijiro Otsuka
May 2018 version
Abstract: Randomized experiments ensure equal opportunities but could generate unequal outcomes by treatment status, which is socially costly. This study demonstrates a sequential intervention to conduct impact evaluation and subsequently to mitigate “experiment-driven” inequality. Specifically, control farmers were initially restricted from exchanging information with treated farmers, who received rice management training, to satisfy the stable unit treatment value assumption. We then encouraged information exchange between farmers one year after the training. We found positive training effects, but performance gaps created by our randomized assignment disappeared over time because of information spillovers and, hence, eventually control farmers also benefitted from our experiment. JEL Code: O12, O13, O31, Q12 Keywords: Inequality, Program evaluation, Randomised experiment, Spillover
*We thank Takeshi Aida, Yutaka Arimoto, Jun Goto, Yoko Kijima, Hisaki Kono, Takashi Kurosaki,
Yuya Kudo, Tomohiro Machikita, Yuko Nakano, Yasuyuki Sawada, Motonori Tomitaka, Takashi
Yamano, Junichi Yamasaki, and seminar participants at the Asian Development Bank, Hitotsubashi
University, JICA Research Institute, Institute of Developing Economies, Kobe University, Kyoto
University, and Sophia University for their valuable comments on earlier versions of this manuscript.
We are also grateful for the financial support by the JICA Research Institute and Ministry of Education,
Culture, Sports, Science and Technology (MEXT), Japan. All views and interpretations expressed in
this manuscript are those of the authors and not necessarily those of the supporting or cooperating
institutions.�
2
1. Introduction
Development, educational, and other social programs have long been chosen and
implemented based on folk wisdom without sufficient scientific evidence. A recent surge
in the use of randomized controlled trials (RCTs) in empirical studies of economics,
particularly of development economics, has substantially contributed to a better
understanding of what works and what does not to improve the welfare of the poor.
Examples include conditional and unconditional cash and asset transfer (Schultz 2004;
Banerjee et al. 2015b; Bastagli et al. 2016; Kabeer and Waddington 2015), microfinance
(Karlan et al. 2014; Banerjee et al. 2015a), preventive health measures (Dupas 2011;
Kremer and Glennerster 2011), educational inputs (Kremer and Holla 2009; Evans and
Popova 2016), and agriculture (de Janvry et al. 2017).
RCTs intentionally, though randomly, create a group of people who receive a
treatment and another group who do not. Such an approach has been justified because it
provides a more credible estimate of the impact of the intervention than other available,
sophisticated econometric methods. Since our preconceptions about the effectiveness of
development practices are often biased, rigorous evidence generated by RCTs helps select
which practice should be scaled-up to a wider range of the population. This holds true
especially when external validity is carefully evaluated from multiple RCTs. Furthermore,
RCTs are justified because everyone in a society cannot receive the treatment
simultaneously, given limited resources, but RCTs, if properly executed, could provide
equal opportunities to all subjects in the targeted population. This ex-ante equality of
opportunities seems to match the sense of fairness of many researchers and plays an
important role in avoiding their ethical dilemmas.
However, RCTs might generate ex-post inequality in outcomes to the extent that the
3
implemented program has positive and significant impacts. Moreover, such inequality
could persist for a long period of time. For example, recent studies on a pioneering
conditional cash transfer (CCT) program in Mexico, Progresa, show that those with more
time exposed to CCTs are significantly better off even more than 10 years after the
program (Kugler and Rojas 2018; Parker and Vogl 2018). Inequality triggered by RCTs
could be eliminated relatively quickly if RCTs are designed to roll out so that everyone
can eventually receive the same treatment. Yet, as is demonstrated by Progresa, which
was precisely designed to do so, this is sometimes imperfect. Alternatively, RCT-triggered
inequality may be mitigated if benefits are allowed to spread from treated to control
groups by such means as social learning or positive spillovers (Miguel and Kremer 2004;
Kremer and Miguel 2007; Oster and Thronton 2012). Although the existence of positive
interpersonal spillover effects is desirable in the real world, it is generally recused in the
RCT setting because it violates the stable unit treatment value assumption (SUTVA),
without which the unbiased impact of implemented programs cannot be estimated. Thus,
spillovers tend to be considered a threat to identification rather than a driver to reduce
otherwise persistent inequality in experimental settings. As a result, researchers who
initiate RCTs often prefer the absence of spillovers and tend to overlook inequality
generated by “researcher-tailored” RCTs or intentionally leave such inequality unattended
to observe longer-term impacts.
In this study, we implement a unique field experiment in collaboration with 295 rice-
growing farmers in Cote d’Ivoire to explore (1) whether management training for rice
cultivation provides the intended positive impacts, such as the adoption of recommended
agronomic practices as well as improved rice yield and profits, and (2) whether any
generated inequality from our intervention could be eliminated later by encouraging
4
treated farmers to disclose new knowledge acquired during the training to control farmers.
The first objective echoes other RCTs designed not to cause spillover effects, and the
second objective deviates from their common practices.
We take up a case study of rice farming in Cote d’Ivoire because, like other West
African countries, rice is one of the major staple foods in this country and its consumption
has increased rapidly in recent years, exceeding the domestic production. The government
has tried to increase rice yield to sustain food security and save foreign exchange reserves.
The adoption rate of modern inputs, such as improved seeds and chemical fertilizer, is
higher in Cote d’Ivoire than in other rice-growing countries in Sub-Saharan Africa (SSA),
owing to past training provided by local governmental and international organizations
such as AfricaRice (formerly known as WARDA [West Africa Rice Development
Association]), whose headquarters was originally located within the country and is now
there again.1 However, several recommended agronomic practices, including straight-
row transplanting, that have proven to boost rice yield in tropical Asia as well as other
SSA countries have not been adopted widely (David and Otsuka 1994; Otsuka and Larson
2013, 2016). There is thus room for management training to improve the performance of
rice production.
Japanese experts offered a short technical training course in 2015 in collaboration
with local extension agencies to a sub-sample of farmers randomly selected from eight
production sites located in two major rice-growing regions, Bellier and Gbeke. To
mitigate noncompliance, such as the participation of ineligible farmers, local counterparts
checked the attendance of participation in field training every time it was implemented.
1 Because of political instability in Cote d’Ivoire, the headquarters of AfricaRice was temporarily moved to Cotonou, Benin from 2005 to 2015.
5
We conducted household surveys three times at the baseline before the training
(January 2015 to May 2015), the mid-line one year after the training (March 2016 to May
2016), and the end-line two years after the training (March 2017 to May 2017). During
the initial experimental phase between the baseline and mid-line survey, treated farmers
were asked not to transmit information taught in the training and control farmers were
requested to refrain from asking treated farmers for agricultural advice. Both treated and
control farmers were convinced that if they met these requirements they could obtain
precise and valuable knowledge about the effectiveness in their settings of the
technological package taught in the training. After the one year of observation, we relaxed
the restriction and turned to promoting spillovers. Using these three-year panel data, we
examine the evolution of both intention-to-treat (ITT) and treatment-on-the-treated (TOT)
effects of the training, with the attendance rate instrumented by the random assignment
of treatment status for the latter.�
A focus on agricultural training and technology is suited to our research purpose
because it has been an area where spillovers likely occur through social learning (Foster
and Rosenzweig 1995; Munshi 2004; Bandiera and Rasul 2006; Conley and Udry 2010).
Most previous studies share the view that training millions of small farmers is a
significant challenge in practice and farmer-to-famer training or social learning, e.g.,
farmer field schools (FFS), is potentially more cost effective in the diffusion of
agricultural technologies (Guo et al. 2015; Emerick et al. 2016; Mekonnen et al. 2018;
Ben Yishay and Mobarak 2018; Nakano et al. 2018).
Nevertheless, there is little consensus as to the relative effectiveness of direct training
by extension workers and learning from peer farmers. To the extent that social learning is
less effective than direct training to counter information failures, inequality generated by
6
an experiment on agricultural training may not easily cease even after information
dissemination from treated to control farmers is encouraged. Feder et al. (2004), Tripp et
al. (2005), and Kondylis et al. (2017) find that directly trained farmers significantly
increase the adoption of the new technology, but their behavior has limited impacts on
other farmers. On the other hand, Krishnan and Patnam (2014) demonstrate that while the
initial impact of extension agents is high, learning from neighbors plays a more important
role than direct training by extension agents in the adoption of the new technology over
time. Similarly, Nakano et al. (2018) show that directly trained farmers perform better
initially, but other farmers can catch up with them later through farmer-to-farmer training.
Finally, Genuis et al. (2014) suggest that both extension services and social learning are
strong determinants of technology adoption, and the effectiveness of each of the two
informational channels is enhanced by the presence of the other. These contradicting
findings suggest that the nature and strength of peer effects is not readily generalizable
and should be evaluated in each specific context (Sacerdote 2014).
Our main findings are summarized as follows. We find that while the adoption rates
of improved rice management practices are generally high even at the baseline, the treated
farmers are more likely to adopt improved practices, such as seed selection, transplanting
in rows, and field leveling, which is required for uniform crop maturity, within a year
after the training. The higher adoption rates of those recommended agronomic practices
lead to improved rice yield and quality as well as increased income per hectare among
treated farmers. Once all farmers are encouraged to exchange information later,
productivity gaps between treated and control farmers narrow sharply. It may seem
possible to interpret this convergence as a sign of short-lived impacts of training where
trained farmers drop new practices and return to the traditional ones. However, we
7
observe that trained farmers continue to adopt the improved agronomic practices two
years after the training, and control farmers follow them. Our detailed network analysis
based on a dyadic regression further reveals that information flow from treated to control
farmers is less active than between control farmers (a reference group) a year after the
training but becomes more active two years after the training. Meanwhile, information
exchange is more active in the first place between treated farmers than between control
counterparts. These results together suggest that farmers followed our guidance not to
exchange agricultural information within the initial experimental phase, which enables us
to rigorously evaluate the short-term impact of training. Yet, once such a restriction is
abolished and information exchange is encouraged, control farmers could successfully
catch up with treated farmers through social learning. These results imply the importance
of social learning not only for the wider diffusion of agricultural technologies but also for
reducing otherwise persistent inequality in experimental settings.�
To the best of our knowledge, this is the first study that attempts to impose SUTVA
in the initial experimental phase but intentionally relaxes it later. Although the existing
RCTs pay due attention to ethical concerns, to date, most studies seem to put too much
emphasis on identifying the efficacy and effectiveness of implemented projects as well as
mechanisms underlying the positive, negative, or negligible impacts. We do not deny the
importance of those studies to fill the significant knowledge gaps. However, it might also
be valuable to build in a mechanism to allow control groups to catch up with treated ones
and rectify inequality generated by an experiment. By reducing the RCTs’ potential social
cost and improving the welfare of the entire sample, the RCT is likely to be more widely
accepted. This study demonstrates that such design is possible by encouraging social
interaction among subjects once the intended intervention proves to provide positive
8
impacts.
The rest of the paper is organized as follows. Section 2 explains the study setting,
sampling framework, and experiment design, and examines the summary statistics of our
sample. Section 3 explains our estimation strategy on the dynamic impact of training and
discusses estimation results. Section 4 conducts a detailed analysis of the information
network and explains the estimation results. Section 5 concludes the study.
2. Survey and Experimental Design
2.1. The Study Area
The study took place in the Bellier and Gbeke regions, near the capital city of
Yamoussoukro, in Cote d’Ivoire. The two regions were selected under a bilateral official
development assistance (ODA) program between the Ivoirian and Japanese governments.
Japanese technical experts were dispatched from 2014 to 2018 to improve domestic rice
production and increase the quantity of marketed rice through the ODA scheme organized
by Japan International Cooperation Agency (JICA). There are a total of 107 production
sites suitable for rice production within those two regions, which are all located in the
lowlands. Some production sites have sufficient access to irrigation water and are able to
cultivate rice twice in a good year. Others are in low-humidity zones (called bas-fonds),
dependent on rainfall. The main rice cultivation season is roughly from July to December.
If irrigated, a second cycle starts around January/February. When water is insufficient,
farmers produce other crops, such as yams and peanuts, or leave the paddy field to fallow.
Since these two regions are agro-climatically more favorable for rice production than
other areas in the country, farmers have received various rice cultivation trainings
provided by international donors, including JICA, World Bank, and AfricaRice, as well
9
as local extension agencies, including Agence Nationale d'Appui au Développement
Rural (ANADER).
Out of 107 sites, two production sites were initially selected for the JICA project in
2014. Thereafter, the target area was expanded every year to cover a total of 26 sites until
2018. This study relies on the data from eight production sites selected in 2015. To choose
our study sites, we closely collaborated with technical experts. Admittedly, selection was
not completely random because technical experts have a target to cover 1,500 hectares of
land within the five-year project period. Thus, the study sites are relatively larger in
operational size than the other remaining sites in the Bellier and Gbeke regions. Since the
impacts of training may potentially vary by agro-ecological and institutional conditions,
we classify all potential production sites into four types depending on the accessibility to
irrigation and the existence of prior rice training: (Type 1) with at least a partial irrigation
facility and experience in training; (Type 2) without irrigation facilities but with
experience in training; (Type 3) with at least a partial irrigation facility but no (or
inadequate) experience in training; and (Type 4) with neither an irrigation facility nor
experience in training. We then selected two of each type of site, generating a sample of
eight production sites in total.
2.2. Sampling Structure and Experimental Design
Prior to the experiment, we had meetings with farmers belonging to agricultural
cooperatives in each selected site. The objective of the meeting was to explain our
implementation plan and obtain consent from farmers. Although technical experts had
experience in rice production training in Cote d’Ivoire and recommended management
practices for lowland rice cultivation were fairly well established in experimental fields,
10
we felt it was important to evaluate the training impact on rice production performance
through an RCT because it is common to observe differences between on-farm and on-
station results. We also wondered whether management practices taught in this most
recent training were ineffective for those who had already received similar training in the
past or those whose productivity was already close to the production possibility frontier.
Based on an agreement with technical experts, we explained our plan to farmers as
follows: (1) We would like to conduct a social experiment to assess the impact of training
and ask farmers to cooperate with us; (2) farmers are randomly grouped into two groups,
with one eligible to receiving the training offered by the JICA experts while the other is
expected to apply the best management practice they had access to; (3) all farmers
including control farmers are provided with necessary inputs, such as improved seeds and
chemical fertilizer, on credit; (4) the experimental phase lasts one year during which
farmers belonging to different groups are expected not to exchange information about
techniques and management practices taught in training. Specifically, we request treated
farmers not to transmit information taught in the training and controlled farmers to refrain
from asking treated farmers for agricultural advice; (5) before and after the experiment,
we will conduct household surveys for impact evaluation; (6) if farmers follow our
guidance and treated farmers do not transmit information on rice production management,
we can obtain reliable estimates of the impact of training; (7) after the impact assessment,
we will share which technology (i.e. conventional practice vs. one taught in training) is
found to be superior; and (8) after the experimental phase, farmers are encouraged to
share information to provide farmer-to-farmer training.
While unequal treatment during the experimental phase could be a source of tension
between treated and control farmers, we attempted to make them feel neither lucky nor
11
unlucky in their treatment status. Rather, we emphasized that once we know which
technology is better, everyone can benefit from such knowledge and that the success of
this social experiment depends crucially on whether farmers exchange information or not
within one year after the training. This sort of explanation seems to ease tensions, and
most farmers understood the purpose of this experiment and showed strong willingness
to cooperate with us.2
After obtaining consent, we collected individual member lists from each agricultural
cooperative. We attempted to randomly select about 100 farmers from each type specified
above. However, since the total number of farmers in Type 3 was only around 50, we
overweighed Type 4 so that the number of farmers with and without past training
experience totalled roughly 200. Out of 414 farmers on the shortlist, 295 households were
found to be active rice producers who cultivated rice at least once in the preceding year,
resulting in 83 farmers in Type 1, 73 farmers in Type 2, 39 farmers in Type 3, and 100
farmers in Type 4. These 295 households constitute the primary sample in this study. We
conducted the baseline survey with those households from January 2015 to April 2015.
The data pertain to household demographic characteristics, accessibility to land and its
tenure status, details of rice production and other income-generating activities, and
household asset holdings.
We then assigned eligibility to participate in the training. One half of sample
households were randomly selected as a treatment group and the other half as a control
group in each site. Randomization was implemented at the farmer level within each site.
Technical experts provided a short classroom training to extension agents of ANADER
2 Some cooperatives voluntarily created rules to prevent control farmers from learning or adopting the agricultural techniques taught in the training.
12
and three key farmers who were selected from each site. Those extension agents and key
farmers in turn offered on-site training to eligible farmers under the supervision of
technical experts. This training consisted of (a) land preparation, including land levelling,
(b) water control, including canal construction and maintenance, (c) seed selection and
incubation, (d) fertilizer and herbicide application, and (e) harvest and post-harvest
management. To mitigate noncompliance, particularly the participation of the control
farmers, local counterparts visited every session of the training and recorded who
participated in it. The on-site training proceeded gradually to meet the actual rice
cultivation cycle and, in total, it was held at least six times from June to November 2015
to cover the key practices. We then conducted the follow-up surveys twice, the first soon
after the training (March 2016 to May 2016) and, the second two years after the training
(March 2017 to May 2017). During the 2015–2016 seasons, however, there was a severe
lack of rainfall. Thus, sample attrition is serious as will be discussed below.
2.3. Descriptive Statistics and Balancing Test
Table 1 presents the composition of our sample plots. We have a total of 424 plots
from 295 households in the baseline survey. Of those, 333 plots were rice planted in the
main season. The number of sample plots in the main season dropped sharply to 193 in
the mid-line and further to 168 in the end-line survey because of the lack of rainfall. Rice
cultivation was difficult in these years even for those with access to irrigation because of
insufficient water. Therefore, second cropping was almost impossible for most farmers.
As a result, the number of plots cultivating rice in the sub-season declined substantially
from the baseline survey. Production sites without irrigation, that is, Types 2 and 4, were
more severely affected by rainfall shortages, and the number of attrited sample farmers
13
from these groups is larger than from the others.
Table 2 shows the balance test on baseline characteristics for the full sample and
sub-sample in the main season. Households with missing values are dropped. We
conducted a t-test of the equality of means between the treated and control farmers and
joint significance F-tests in columns (3) and (6).
On average, households are large (about nine persons), headed by a male in the mid-
40s with minimal or no formal education. The average plot size is relatively small—
approximately 0.5 hectare. Although the treatment status was randomized, the difference
in the plot size between treatment and control farmers is statistically significant for both
full and sub samples. Most land was operated under owner cultivation. If rented, it was
generally under a fixed-rent contract. Attendance rates at the training was about 42%
among treated farmers, while it was almost null (only two cases) among control farmers,
indicating that almost all control farmers adhered to our request and did not participate in
the training. We conducted the joint significant test except for the attendance rate,
demonstrating that we reject the zero-null hypothesis for the full sample, while we fail to
reject it for the sub-sample in the main season. Given that baseline covariates are balanced
only in the main season and that the second crop likely involves self-selection, we focus
on the main season crop in the subsequent analysis.
Table 3 compares the baseline characteristics of attrited and non-attrited samples
with a t-test of the equality of the mean between the two and the associated joint
significance F-test. The attrition rate differs notably by the accessibility of irrigation, and
the share of the sample from irrigation sites (i.e., Types 1 and 3) is significantly larger
among non-attrited samples. Furthermore, most key observable characteristics are
statistically significantly different between attrited and non-attrited samples: on average,
14
attrited samples are more likely to be female-headed with less education, larger in
household size but smaller in plot size, and more likely to own rice plots. The joint
significance test shows that the zero-null hypothesis is strictly rejected, implying that
attrition is non-random. This non-random sample attrition is a potential threat to our
statistical inference, which should be addressed in the econometric analysis.
Before discussing our empirical strategy in detail, Table 4 presents the changes in
outcome variables of interest regarding rice management practices and productivity of
non-attrited samples over time. We again show the results of t- and F-tests for treated and
control plots. In addition, columns (10) and (11) present an unconditional difference-in-
differences (DID) regression estimate of the treatment effect (i.e. the difference in the
time trend between treated and control plots).
The adoption of recommended management practices was generally high even in the
baseline (Panel A). Because of its proximity to AfricaRice, the adoption of the modern
variety of rice was complete and uptake rates reached 100%.3 The use of chemical
fertilizers was also remarkably high by SSA standards: on average, more than 200 kg/ha
of fertilizer, such as NPK and UREA, were applied. In addition to these external inputs,
the adoption of improved agronomic practices helps boost rice yield in SSA (Otsuka and
Larson 2013, 2016). Water canal/drainage construction and maintenance are important to
manage water levels in rice fields during the growth period, while levelling is crucial to
reduce the amount of water wasted by uneven pockets and to promote even growth of rice
plants. Straight-row planting can be adopted to facilitate other complementary
management practices such as hand or rotary weeding and even the application of
3 The vast majority of farmers used WITA-9, a high-yielding variety that is tolerant to rice yellow mottle virus and iron toxicity with a maturity period of about 110 days.
15
fertilizers, herbicides, or insecticides. In our sample, 75–90% of plots had levelled fields
and constructed/repaired water canal/drainages in the baseline. Most sample farmers
selected better seeds by water or winnowing, whereas transplanting in row was less
common.
Panel B shows the rice productivity and profitability of sample plots. Gross
production value per hectare is computed by multiplying the rice yield (1000 kg/ha) with
the price received (kg/CFAF).4 Rice income per hectare is equal to the gross production
value minus paid-out costs, including land rent, irrigation fees, costs of purchased
chemicals, and machinery rental, divided by the plot size. Profits per hectare are equal to
rice income minus imputed family labor costs, divided by the plot size. To impute family
labor costs, we used the typical prevailing hired wage rate for transplanting in each village.
The average yield exceeds 3.4 tons/ha which is significantly higher than the average of
other countries in SSA of just above 2 tons/ha (Otsuka and Larson 2016). The average
gross output value, rice income, and profits per hectare were about 600 thousand CFAF
(or approximately 1,065 USD), 405 thousand CFAF (or 719 USD), and 320 thousand
CFAF (or 568 USD), respectively.
The table also shows that while there is no statistically significant difference in the
baseline adoption rate of recommended practices, treated farmers are more likely to adopt
levelling, canal/drainage construction/repairs, and transplanting in row at the time of the
mid-line survey. Although they tend to adopt those practices more than control farmers in
the end-line, the unconditional DID estimate shows that the incremental adoption rate
between the mid- and end-line is higher for control farmers. On the other hand, all
outcome variables but rice yield are not significantly different between the treated and
4 1 USD is equivalent to 563 CFAF as of January 2015.
16
control samples in the baseline, and no outcomes are significantly different in the mid-
line and end-line surveys. However, the unconditional DID estimates show that treatment
plots increase rice yield and revenue between the baseline and mid-line more than control
plots, while the reverse was true between the mid-line and end-line surveys.
These results suggest that the treated farmers improved their rice management
practices and performed better in the first year after the training when information
exchange was restricted, but control farmers caught up with treated farmers once
information sharing was encouraged presumably because of spillover effects.
3. Dynamic Impacts of Training
3.1. Estimation Strategy
To identify the causal relationships between the provision of training and outcomes
of interest, we estimate intention-to-treat (ITT) and treatment-on-the-treated (TOT)
effects. We first examine the average impacts of all production sites (i.e., Types 1–4),
allowing the impacts to vary across time. We are particularly interested in whether the
training brings intended positive impacts in the first year after the training with the gap
generated by the experiment decreases over time through spillovers in the next year.
Following McKenzie (2012), we employ an analysis of covariance (ANCOVA) model in
the form of:
!"#$ = &' + )!"#' + &*+$ + &,-"# + &.(+$ × -"#) + 2"#'3 + 4# + 5"#$ (1),
where !"#$ and !"#' are the post- and pre-treatment outcome variables of plot i in
production site j at time t (i.e., either mid-line or end-line) and time 0 (i.e., baseline);+$
is a dummy variable for the end-line data; -"# is a dummy variable equal to one if a
household is eligible to participate in the training (ITT estimate) or a continuous variable
17
for the attendance rate of training, instrumented by the treatment status (TOT estimate)5;
2"#' is a set of baseline control variables; 4# is the time-invariant fixed effect at the
production site; and 5"#$ is the unobserved error term. The parameters of interest are &,
and &.. The former captures the short-term impacts of training under the imposition of
the SUTVA, while the latter represents the mixture of the longer-term training impacts
and spillover effects when the SUTVA is relaxed. We note that the pure training impact is
estimable only in the short-term.
As outcome variables, we focus on the use of chemical fertiliser (kg/ha), the adoption
of seed selection by water or winnowing (=1), levelling (=1), canal/drainage
construction/repairing (=1), and transplanting in row (=1) as well as rice yield (ton/ha),
gross output value (‘000CFAF), rice income (‘000CFAF), and rice profit (‘000CFAF) per
hectare. When the outcome is binary, we apply a liner probability model. As baseline
control variables, we include household size, household head’s characteristics (including
age, gender, and years of education), plot characteristics (including parcel size and tenure
status dummies), and the logged value of household assets at the baseline survey. We
cluster all standard errors within production sites.
The random assignment of treatment status should make the treatment and control
groups similar in expectation. Therefore, including controls in regressors and/or applying
the ANCOVA model would not affect the consistency of the estimated treatment effects.
However, the inclusion of additional controls is expected to lend greater credibility to
internal validity of the estimates when some baseline imbalance exists. Thus, we prefer
and present the results with baseline controls. Still, the estimated parameters may be
5 Strictly speaking, this is the local average treatment effect (LATE). However, because almost no control farmers attended the on-site training, our estimate can be virtually considered TOT (Angrist and Pische 2008).
18
biased due to non-random sample attrition. To adjust for that, we use the inverse-
probability weighting method, suggested by Wooldridge (2010). Specifically, we run the
probit regression to compute the predicted probability of non-attrition and use the inverse
of it as weights in the main equation. This first-stage probit regression result is presented
in Appendix 1.
3.2. Estimation results
Table 5 shows the estimation results for the dynamic impacts of management training
on rice productivity and profitability. For the sake of brevity, coefficients on control
variables are suppressed.
It is clear that training has positive and significant impacts on rice productivity by
the mid-line, with the rice yield increasing by 0.75 ton/ha, gross output value per hectare
by 140 thousand CFAF, and rice income per hectare by 103 thousand CFAF. These
improvements correspond to 20%, 24%, and 29% of control means, respectively,
suggesting that management training was effective in our context.6 This improvement in
productivity, however, does not lead to an increase in profits. As we will see, this is
presumably because trained farmers test a larger number of improved management
practices than control farmers, who require more family labor inputs. Qualitatively
similar results are observed for TOT estimates. The fact that we see quantitatively larger
magnitudes of impacts in TOT than ITT estimates suggests that actual training
participation rather than its simple eligibility is important to improve production
performance.
6 According to experienced agricultural experts, impacts of recommended management practices on rice productivity are generally larger when there is sufficient water. Thus, our estimates could be considered the lower bound of the impacts that would be realized in a year with normal rainfall.
19
Notably, the coefficient estimates on the interaction term are negative and significant
for rice yield and gross output value per hectare. This indicates that the improvement of
performance among treatment groups from the mid- to end-line is lower than control
groups. The Wald test shows that we cannot reject the null hypothesis that the total
training effect is zero in most specifications, implying that treated farmers are no better
than control farmers by the end-line. We can interpret this negative interaction term, &.,
as reflecting either the short-lived training effects or the existence of spillover effects. If
training impacts do not last long, however, we would observe some signals, such as the
declining adoption rate of improved management practices among treated farmers. We
did not observe clear disadoption patterns in Table 4. Thus, this finding seems consistent
with the operation of a mechanism wherein control farmers improve their performance
by learning from treated farmers after the SUTVA is relaxed.
Table 6, which shows estimated impacts of training on the adoption of improved
agronomic practices, also provides supportive evidence of spillovers. When information
exchange between treated and control farmers was restricted during the year after the
training, the positive training impact on the adoption of improved management practices,
such as levelling, canal/drainage construction/repairing, and straight-row transplanting is
observed among treated farmers (ITT estimate) and training participants (TOT estimate).
However, once the restriction was lifted two years after the training, control farmers
successfully caught up with treated farmers in the adoption of recommended practices, as
reflected in the negative and significant coefficients on the interaction term, &..7 The
Wald tests also confirm that in most outcomes we fail to reject the hypothesis of zero
7 Since the same amount of fertilizer is provided to both treatment and control groups in the experimental phase, it seems plausible to observe insignificant effects of training on this outcome.
20
training impact in the longer term.
Taken together, we confirm that training has positive impacts in the short-term not
only on the adoption of improved rice management practices but also on rice productivity.
Our further intervention encouraging farmers to spread information contributes to
reducing the generated gap.8 In order to ascertain whether this result reflects spillovers,
we will examine in more detail whether social networks actually mediate the information
spillover in Section 4.9
3.3. Heterogeneous treatment effects
Before moving on to the detailed network analysis, we investigate the heterogeneous
treatment effects by the type of production site. The purpose of this analysis is to
determine whether this training had negligible impacts on those who had already received
similar training in the past or those whose production was already close to the production
possibility frontier. Since the introduction of multiple interaction terms and multiple
endogenous variables makes interpretation complex, we estimate time-invariant ITT
effects by incorporating the interaction term only between the treatment and type
dummies. We thus ignore the differentiated dynamic impact of the training across
production sites over time and their corresponding TOT estimates.
The estimated results in Table 7 show that the training had positive impacts on the
8 Although the number of our outcome variables is not so large, one may wonder if we find false positives because of testing multiple hypotheses. To address this concern, we compute false discovery rate sharpened q-values corrected multiple testing, following the Benjamini-Kreieger-Yekutieli method (Bendamini et al. 2006). All outcome variables that show statistically significant effects in Tables 6 and 7 remain significant at 10% or lower. 9 If spillovers exist, the average performance of control groups would improve over time, which could be reflected in &* (the end-line dummy)>0. &* is positive for most outcome variables, and statistically significant for the adoption of canal/drainage construction/repairing and straight-row planting for TOT estimation, further supporting our interpretation in favor of the existence of spillovers.
21
adoption of recommended rice management practices and rice productivity, including rice
yield, the use of chemical fertilizer per hectare, and the application of levelling. Contrary
to our expectation, statistically significant impacts prevail, especially for Type 1, where
rice cultivation environments are most favorable, and farmers have sufficient experience
with training in the past. Those who belong to Type 4, for which production environments
are least favorable and training experience is scant, do not learn much from training, as
indicated by the negative and significant coefficient estimates on the treatment dummy
interacting with the type dummy for most specifications. This might show the low
expected returns of improved rice production methods in rain-fed areas where water
control is difficult, which was also the case in Asia (David and Otsuka 1994).
4. Spillover Effects
4.1. Information Network Analysis
Having shown the treatment effects across time and production sites, we now
examine whether social networks actually mediate information spillover from treated to
control farmers, using the detailed learning link data.
A fundamental empirical challenge on this topic is how to correctly specify one’s
social network. Asking respondents about their social network by arbitrarily setting a cap
on the number of links may result in truncation bias, while asking an open-ended question
tends to capture only the strong links, ignoring the weaker ones (see, for example,
Maertens and Barrett [2013] for a thorough discussion of potential bias in empirically
eliciting the true social network structure). To address this concern, we exploit a “random
matching within sample” technique to elicit social networks, following, among others,
Conley and Udry (2010), Maertens and Barrett (2013), and Mekonnen et al. (2018). More
22
specifically, we match each sample respondent with six other survey respondents
randomly drawn from the sample in the same production sites and ask details of the
(non)existence of information exchange about agronomic practices between sample
farmers. To examine the differential roles played by treatment and control peers, we select
three matches from treated farmers and the other three from control farmers. To capture
changes in the network of interactions over time, we collected the learning link data in
both the mid- and end-line surveys. As Santos and Barrett (2008) demonstrate, the
random-matching-within-sample method recovers the underlying social network
structure more reliably than other available methods, such as a network-within-sample
method in which each respondent is asked about his/her link to every other household in
the sample.
We then run a dyadic regression for those who know their matches to characterize
the flow of information about management practices across farmers over time. 10
Formally, let 9"#$ be equal to one if a respondent farmer i asks farmer j (conditional on i
knows j) for advice on agronomic practice, such as land preparation, transplanting, and
fertilizer application, at time t.11 We explore the correlates of learning links by including
attributes of a household i and j as:
9"#$ =δ + γT= + α*-"#*+α,-"#, + α.-"#.+β*(-"#* × T=) + β,(-"#, × T=) + β.(-"#. × T=)
+@2" + 2#AB + @2" − 2#AD +E"#F + G + H"#$ (2),
where -"#* , -"#, , and -"#. are a combination of the treatment status of households i and j
with [treated, treated], [treated, control], and [control, treated]. The remaining
10 Using the full sample, including nonacquaintance pairs, does not alter our main findings. 11 In the questionnaire, we asked respondents whether he/she asked advice on rice management practice from a specified person by the time of the survey. 9"#$ is one if the answer to this question is yes.
23
combination [control, control] is a reference group; T is a binary indicator for the end-
line survey; 2" and 2# denote a vector of baseline controls for farmers i and j
characteristics, respectively12;E"# describes a dummy equal to one if the gender of both
farmers is the same; G is the production site fixed effect; and H"# is a random
disturbance. Following Attanasio et al. (2012), standard errors are clustered at the
production site level to allow for possible correlations not only within dyadic pairs but
also across all dyads in the same location.
Table 8 presents estimated results by a linear probability model. The coefficient
estimate on the [treated, treated] dummy, α*, is positive and statistically significant, but
its interaction term with the end-line data dummy, β*, is statistically insignificant. This
indicates that information exchange amongst treated farmers is more active than control
counterparts at the same production site, and its tendency does not systematically change
over time. On the other hand, consistent with our expectation, we observe a negative and
significant coefficient on the [control, treated] dummy, α., and a positive and significant
coefficient on the interaction term with the end-line data dummy, β. . These results
illustrate that either controlled farmers refrained from asking agricultural advice from
treated farmers or the latter refrained from disclosing the management information to the
former in the first year after the training,13 but they are eager and active in doing so in
the second year after the training. This strongly indicates that impact evaluation in the
initial phase was less likely to be undermined by spillovers, supporting our claim that the
recommended practices were more productive. It also supports our main finding that there
12 If 9"# is bidirectional (i.e., 9"# = 9#"), &2"# = &2#" should be imposed: In such a case,|2" − 2#| instead of @2" − 2#A is more relevant as regressors (Fafchamps and Gubert 2007). 13 This is also consistent with the fact that some cooperatives created their own rules to keep control farmers from learning the management practices taught in the training during the first year.
24
were information spillovers after the relaxation of the SUTVA in the two years after the
training, which would facilitate control farmers to improve their rice management
practices and performance through social learning.
4.2. Extension to the linear-in-mean model
While our analysis so far supports the existence of social learning, one may wonder
whether respondents’ self-report about information exchange patterns is biased and
simply reflects their willingness to satisfy researches’ expectation. Although we cannot
directly address such concerns, if social learning actually plays a role, we might observe
the influences of peer behaviour and performance on one’s own. As a robustness check
to verify this possibility, we employ the linear-in-mean model. Exactly identifying peer
effects is, however, complicated because of reflection problems. As discussed by Manski
(1993), individuals would behave similarly not only through social learning (called the
“endogenous effect”), but also because they have similar characteristics (called the
“exogenous effect”) or they face similar institutional environments (called the “correlated
effect”). To disentangle endogenous social effects from other confounders, we add several
variables to Equation (1).
First, based on the network data elicited in the random-matching-within-sample
method, we take the average values of baseline observable characteristics in i’s
information network, regardless of whether network peers are treated or control farmers,
and include them into regressors. This serves to control for exogenous effects along with
the location fixed effects that control for unobservable correlated effects. Second, to
explore whether the peers’ average behavior and performance directly affect farmer i’s
performance, we also include the average productivity or technology adoption in i’s
25
network, regardless of their treatment status. Since the number of peer adopters in i’s
information network is used for the numerator of technology adoption variables, this
mimics an identification strategy undertaken by Bandiera and Rasul (2006). 14 , 15
Following Mekonnen et al. (2018), we use lagged rather than contemporaneous values of
mean group performance or behavior in recognition that information on agricultural
technology cannot be diffused quickly. Third, as another variable to capture endogenous
peer effects especially mediated by treated farmers, we add the share of treated farmers
in i’s information network along with the network size (i.e., max six). This is akin to the
methodology used by Kremer and Miguel (2007) and Oster and Thronton (2012). The
original intuition behind this method is that once we control for the network size (which
could be potentially endogenous),16 the share of network peers in the treatment group is
random because of the randomized experiment. This exogenous variation can be then
used to identify peer effects.
Note that the average peer performance and the share of treated farmers in i’s network
above are expected to reflect different channels of peer effects: the former may partly
capture learning by direct observation even without mouth-to-mouth communication,
while the latter may partly capture knowledge transmission from treated farmers even
when treated farmers do not actually adopt some technologies due to certain constraints.
We restrict the observation of this analysis to the end-line year as it reflects a normal
14 Unlike Bandiera and Rasul (2006), however, we created the variable of peer adopters based not on respondent’s perception (i.e., proxy-report), but on peer’s own self-reports in the interview. See Hogset and Barrett (2010) for a discussion on how proxy-report tends to contain a measurement error and possibly biases estimation results. 15 Following Bandiera and Rasul (2006), we have included the quadratic term of the average outcome in i’s network as well, but we did not find significant non-linear relationships. 16 In the random matching within sample method, the network size should not be interpreted literally, but rather as a proxy for one’s social connectedness where the more random matches a household knows, the larger will be their true social network (Murendo et al. 2018).
26
condition without the prohibition of information exchange, in which spillovers are more
likely to take place. We also restrict the outcome variables to rice yield, gross output value
per hectare, the adoption of field levelling, canal/drainage construction/reparing, and
straight-row planting for which strong information spillovers from treated to control
farmers seem to exist as observed in Table 6.
Although we attempt to minimize concerns about spurious correlation by controlling
exogenous and correlated effects, we are aware of a potential endogeneity issue in this
exercise. For example, control farmers who are more motivated, if all else held constant,
may be more willing to establish information links with treated farmers who know the
new technique or with peers who actually adopt it. Given the possibility that interventions
can alter the underlying network structure (Comola and Prina 2017; Advani and Malde
2018), we admit that our constructed variables to capture social learning/endogenous peer
effects are not free from endogeneity concerns.17
With this caveat in mind, the estimated results in Table 9 show that the share of
treated farmers significantly positively affect one’s own behavior and performance for
gross output value per hectare, levelling, and straight-row planting. We also observe a
positive effect of the average performance of network members in most specifications,
although they are statistically insignificant except for levelling, perhaps due partly to the
low statistical power and partly to difficulties in mimicking new technologies without
learning through deep communication. While these results should be interpreted with
caution to avoid strong causal inference, it seems to be no exaggeration to argue that the
results provide further suggestive evidence on the existence of spillovers, especially
17 Without a sufficient number of instruments, it seems technically infeasible to overcome this potential endogeneity issue.
27
mediated through treated farmers.
5. Conclusions
The use of RCTs in empirical studies particularly in development economics has been
growing rapidly and providing a substantial amount of new knowledge on real-world
practices. RCTs, however, could be a potential source of inequality in outcomes to the
extent that they offer positive benefits to the treated individuals, giving rise to ethical
concerns. By facilitating spillovers from treated to control individuals, we could mitigate
such inequality and improve social welfare; however, it comes at a cost in that spillover
could be a threat to accurate impact assessment because it violates the SUTVA.
This study executes a novel RCT to examine whether rice management training has
the intended positive impacts on the adoption of recommended practices and productivity
in the short term as well as whether any performance gap between treated and control
farmers diminishes over time by facilitating information spillovers from the former to the
latter in the subsequent period. We found the positive and significant short-term effects
of the training, which widen the gap in yield by 20%, gross output value per hectare by
24%, and the adoption rates of selected rice management practices between treated and
control farmers. However, during the technology dissemination process, control farmers
improved their performance, and, as a result, the gap between treated and control farmers
becomes virtually zero in the longer term. Our detailed analysis of learning link data
shows supportive evidence of the existence of information spillovers. Indeed, control
farmers are less likely to ask treated farmers for advice on rice management practices
when we asked them not to do so and the two groups of farmers are more likely to
exchange new knowledge after the restriction is lifted. Thus, the benefit of the direct
28
training by extension agencies has been effectively spread from trained to non-trained
farmers through indirect farmer-to-farmer information transmission over time.
More often than not, experimental studies create inequality in the community that is
never addressed. This inequality may be socially costly and sometimes unacceptable to
potential beneficiaries and practitioners. In those cases, researchers would lose
opportunities to deeply understand the causes and resolutions of social problems, while
communities and practitioners might also miss opportunities to adopt effective programs.
By contrast, the present study has demonstrated a unique research design, in which we
initially run the RCT to establish the impact of an intervention, and we next facilitate the
dissemination of the useful information to the rest of the community by providing subjects
with opportunities for social learning.
We expect that the benefit of this new research design is particularly pronounced in
development programs whose benefits can be shared by local community members.
Examples include training in the use of agricultural technology, entrepreneurial
management, and health care, wherein local community members can simultaneously
benefit from new knowledge without jeopardizing others’ use and without incurring
substantial training cost.
One specific policy implication drawn from our study is that since knowledge of
agronomic management practices is likely to be a local public good, it should be
disseminated at least initially by the public sector through such means as agricultural
training. Otherwise, efforts to improve technology would be suboptimum because of non-
excludability of agricultural technology. In all likelihood, it is a mistake to assume that
profit-oriented private enterprises provide technology information, e.g. through contract
farming, unless information dissemination through social networks is not feasible. Public
29
extension agents can save resources by offering the extension services only to selected
farmers in the community who would then offer a technical training to neighboring
farmers. Although our experiment relies on random assignments without any incentive
scheme, how to best select treated nodes and whether incentives should be given to them
in order to maximize diffusion and efficiency may be important questions for future
research (Beaman et al. 2015; Emerick et al. 2016; Kondylis et al. 2017: Barrett et al.
2018; BenYishay and Mobarak 2018).
Another more general policy implication to be stressed is the importance of rigorous
impact evaluation using RCTs combined with other methodologies to identify useful
knowledge for local people. This is crucial to avoid the persistence of traditional inferior
practices or the introduction of inappropriate knowledge, which can take place if a social
experiment is dispensed with. Herein lies a new important opportunity for economists and
experts in agricultural, management, and health sciences to collaborate on for the sake of
improving the well-being of a group of people.
30
Reference Advani, Arun, and Bansi Malde. 2017. “Credibly Identifying Social Effects: Accounting
for Network Formation and Measurement Error.” Journal of Economic Survey (forthcoming).
Angrist, Joshua D, and Jorn-Steffen Pischke. 2008. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press: Princeton.
Attanasio, Orazio, Abigail Barr, Juan C Cardenas, Garance Genicot, and Costas Meghir. 2012. “Risk Pooling, Risk Preferences, and Social Networks.” American Economic Journal: Applied Economics 4 (2): 134–167.
Bandiera, Oriana, and Imran Rasul. 2006. “Social Networks and Technology Adoption in Northern Mozambique.” Economic Journal 116(514): 869–902.
Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan. 2015a. “The Miracle of Microfinance? Evidence from a Randomized Evaluation.” American Economic Journal: Applied Economics 7 (1): 22-53.
Banerjee, Abhijit, Esther Duflo, Nathanael Goldberg, Dean Karlan, Robert Osei, William Parienté, Jeremy Shapiro, Bram Thuysbaert, and Christopher Udry. 2015b. “A Multifaceted Program Causes Lasting Progress for the Very Poor: Evidence from Six Countries.” Science 348 (6236): 1260799.
Barrett, Christopher, Paul Christian, and Abebe Shimeles. 2018. “The Processes of Structural Transformation of African Agriculture and Rural Spaces.” World Development 105: 283-285.
Bastagli, Francesca, Jessica Hagen-Zanker, Luke Harman, Valentina Barca, Georgina Sturge, and Tanja Schmidt. 2016. Cash Transfers: What Does The Evidence Say? A Rigorous Review of Programme Impact and of the Role of Design and Implementation Features. ODI: London.
Beaman, Lori, Ariel Benyishay, Jeremy Magruder, and Ahmed M. Mobarak. 2015. “Can Network Theory-Based Targeting Increase Technology Adoption ?” Working Paper. Northwestern University.
Benjamini, Yoav, Abba M. Krieger, and Daniel Yekutieli. 2006. “Adaptive Linear Step-up Procedures that Control the False Discovery Rate.” Biometrika 93 (3): 491-507.
BenYishay, Ariel, and Ahmed M. Mobarak. 2018. “Social Learning and Incentive for Experimentation and Communication.” Review of Economic Studies (conditionally accepted).
Comola, Margherita, and Silvia Prima. 2017. “Treatment Effect Accounting for Network Changes: Evidence from a Randomized Intervention.” Mimeo.
Conley, Timothy G, and Christopher R. Udry. 2010. “Learning about a New Technology:
31
Pineapple in Ghana.” American Economic Review 100 (1):35–69. David, Christina C, and Keijiro Otsuka. 1994. Modern Rice Technology and Income
Distribution in Asia. Boulder, USA: Lynne Rienner Publishers. de Janvry, Alain, Elisabeth Sadoulet, and Tavneet Suri. (2017). “Field Experiments in
Developing Country Agriculture.” In Handbook of Economic Field Experiments, Vol. 2, edited by A. Banerjee, and E. Duflo, 427–466Amsterdam: Elsevier.
Dupas, Pascalin. 2011. “Health Behavior in Developing Countries.” Annual Review of Economics 3:425-449.
Emerick, Kyle, and Manzoor H. Dar. 2017. “Enhancing the Diffusion of Information about Agricultural Technology.” Unpublished manuscript.
Evans, David, and Anna Popova. 2016. “Cost-Effectiveness Analysis in Development: Accounting for Local Costs and Noisy Impacts.” World Development 77: 262-276.
Fafchamps, Marcel, and Flore Gubert. 2007. “The Formation of Risk Sharing Networks.” Journal of Development Economics 83 (2): 326–350.
Feder, Gershon, Rinku Murgai, and Jaime B. Quizon. 2004. “The Acquisition and Diffusion of Knowledge: The Case of Pest Management Training in Farmer Field Schools, Indonesia.” Journal of Agricultural Economics 55 (2):221–43.
Foster, Andrew. D, and Mark R. Rosenzweig. 1995. “Learning by Doing and Learning from Others: Human Capital and Technical Change in Agriculture.” Journal of Political Economy 103 (6): 1176–1209.
Hogset, Heidi, and Christopher. B. Barrett. 2010. “Social Learning, Social Influence, and Projection Bias: A Caution on Inferences Based on Proxy Reporting of Peer Behavior.” Economic Development and Cultural Change 58 (3): 563–589.
Genius, Margarita, Phoebe Koundouri, Céline Nauges, and Vangelis Tzouvelekas. 2014. “Information Transmission in Irrigation Technology Adoption and Diffusion: Social Learning, Extension Services, and Spatial Effects.” American Journal of Agricultural Economics 96 (1):328–44.
Guo, Mingliang, Xiangping Jia, Jikun Huang, Krishna B. Kumar, and Nicholas E. Burger. 2015. “Farmer Field School and Farmer Knowledge Acquisition in Rice Production: Experimental Evaluation in China.” Agriculture, Ecosystems and Environment 209:100–107.
Kabeer, Nalia, and Hugh Waddington. 2015. “Economic Impacts of Conditional Cash Transfer Programmes: A Systematic Review and Meta-analysis.” Journal of Development Effectiveness 7 (3): 290-303.
Karlan, Dean, Aishwarya L. Ratan, and Jonathan Zinman. 2014. “Savings by and for the Poor: A Research Review and Agenda.” Review of Income and Wealth 60 (1): 36–
32
78. Kondylis, Florence, Valerie Mueller, and Jessica Zhu. 2017. “Seeing Is Believing?
Evidence from an Extension Network Experiment.” Journal of Development Economics 125:1–20.
Kremer, Michael, and Rachel Glennerster. 2011. “Improving Health in Developing Countries: Evidence from Randomized Evaluations.” In Handbook of Health Economics Volume 2, edited by Mark V. Pauly, Thomas G. Mcguire and Pedro P. Barros, 201-315.: Elsevier.
Kremer, Michael, and Alaka Holla. 2009. “Improving Education in the Developing World: What Have We Learned from Randomized Evaluations?” Annual Review of Economics 1: 513-542.
Kremer, Michael, and Edward Miguel. 2007. “The Illusion of Sustainability.” Quarterly Journal of Economics 132 (4): 1007-1065.
Krishnan, Pramila, and Manasa Patnam. 2014. “Neighbors and Extension Agents in Ethiopia: Who Matters More for Technology Adoption?” American Journal of Agricultural Economics 96 (1):308–27.
Kugler, Adriana D, and Ingrid Rojas. 2018. “Do CCTs Improve Employment and Earnings in the Very Long-Term? Evidence from Mexic).” NBER Working Paper No. w24248. Available at SSRN: https://ssrn.com/abstract=3112035
Maertens, Annemie, and Christopher B. Barrett. 2013. “Measuring Social Networks’ Effects on Agricultural Technology Adoption.” American Journal of Agricultural Economics 95 (2):353–59.
Manski, Charles F. 1993. “Identification of Endogenous Social Effects: The Reflection Problem.” Review of Economics Studies 60 (3): 531-542.
McKenzie, D. (2012). “Beyond Baseline and Follow-up: The Case for More T in Experiments.” Journal of Development Economics 99: 210-221.
Mekonnen, Daniel Ayalew, Nicolas Gerber, and Julia Anna Matz. 2018. “Gendered Social Networks, Agricultural Innovations, and Farm Productivity in Ethiopia.” World Development 105: 321-335.
Miguel, Edward, and Michael Kremer. 2004. “Worms: Identifying Impacts on Education and Health in the Presence of Treatment Externalities.” Econometrica 72 (1): 159-217.
Murendo, Conrad, Meike Wollni, Alan De Brauw, and Nicholas Mugabi. 2018. “Social Network Effects on Mobile Money Adoption in Uganda.” Journal of Development Studies 54 (2): 327-342.
Munshi, Kaivan. 2004. “Social Learning in a Heterogeneous Population: Technology
33
Diffusion in the Indian Green Revolution.” Journal of Development Economics 73 (1): 185–213.
Nakano, Yuko, Takuji W. Tsusaka, Takeshi Aida, and Valerien O. Pede. 2018. “Is Farmer-to-Farmer Extension Effective? The Impact of Training on Technology Adoption and Rice Farming Productivity in Tanzania.” World Development 105: 336-351.
Otsuka, Keijiro, and Donald F. Larson, eds. 2013. An African Green Revolution: Finding Ways to Boost Productivity on Small Farms. Dordrecht, Netherlands: Springer.
———. 2016. In Pursuit of an African Green Revolution: Views from Rice and Maize Farmers’ Fields. Dordrecht, Netherlands: Springer.
Oster, Emily, and Rebecca Thronton. 2012. “Determinants of Technology Adoption: Private Value and Peer Effects in Menstrual Cup Take-Up.” Journal of the European Economic Association 10 (6): 1263-1293.
Parker, Susan W, and Tom Vogl. 2018. “Do Conditional Cash Transfers Improve Economic Outcomes in the Next Generation? Evidence from Mexico.” NBER Working Paper No. w24303. Available at SSRN: http://www.nber.org/papers/w24303.pdf
Sacerdote, Bruce. 2014. “Experimental and Quasi-Experimental Analysis of Peer Effects: Two Steps Forward?” Annual Review of Economics 6: 253-272.
Santos, Paul, and Christopher B. Barrett. 2008. “What Do We Learn about Social Networks When We Only Sample Individuals? Not Much.” Unpublished manuscript.
Schultz, T. Paul. 2004. “School subsidies for the poor: evaluating the Mexican Progresa poverty program.” Journal of Development Economics 74 (1): 199-250.
Tripp, Robert, Mahinda Wijeratne, and V.Hiroshini Piyadasa. 2005. “What Should We Expect from Farmer Field Schools? A Sri Lanka Case Study.” World Development 33 (10):1705–1720.
Wooldridge, Jeffery M. 2010. Econometric Analysis of Cross Section and Panel Data. 2nd Edition. Cambridge: The MIT Press.
34
Table 1. Sample Structure by Season, Year, and Type of Production Sites
YearMain
seasonSub
seasonType 1 Type 2 Type 3 Type 4 Total
2015 (baseline) 333 91 170 76 53 125 424
2016 (midline) 193 7 87 23 41 49 200
2017 (endline) 168 46 83 31 54 46 214
Total 694 144 340 130 148 220 838
35
Table 2. Baseline Balance of Sample Plots by Treatment Status
Standard deviations are in brackets. ***<0.01; **<0.05; *<0.1
Treated Control Mean difference Treated Control Mean difference(1) (2) (3) (4) (5) (6)
Attendance rate (=1) 0.421 0.002 0.420*** 0.398 0.002 0.396***[0.023] [0.001] [0.026] [0.001]
Head's age (years) 43.861 43.333 0.527 44.175 44.030 0.145[0.949] [0.832] [1.068] [0.947]
Head's education (years) 3.314 2.900 0.414 3.151 2.821 0.330[0.271] [0.269] [0.311] [0.302]
Head is male (=1) 0.889 0.886 0.004 0.869 0.875 -0.006[0.022] [0.022] [0.027] [0.026]
HH size 9.115 8.724 0.392 9.100 8.863 0.237[0.307] [0.310] [0.353] [0.359]
Plot size (ha) 0.587 0.476 0.110*** 0.571 0.472 0.099*[0.038] [0.020] [0.046] [0.023]
Owner (=1) 0.731 0.776 -0.045 0.756 0.774 -0.018[0.031] [0.029] [0.034] [0.032]
Leaseholder (=1) 0.221 0.162 0.059 0.188 0.161 0.027[0.029] [0.025] [0.031] [0.028]
Sharecropper (=1) 0.024 0.014 0.010 0.025 0.018 0.007[0.011] [0.008] [0.012] [0.010]
Others (=1) 0.019 0.048 -0.028 0.025 0.048 -0.023[0.010] [0.015] [0.012] [0.016]
Log asset value 4.674 4.703 -0.029 4.561 4.637 -0.075[0.099] [0.085] [0.113] [0.091]
F-test of joint significance 1.680* 0.912Number of Observations 208 210 160 168
Full sample Subsample in main season
36
Table 3. Baseline Balance of Sample Plots by Attrition Status
Standard deviations are in brackets. ***<0.01; **<0.05; *<0.1
AttritedNon-
attritedMean
difference(1) (2) (3)
Treatment (=1) 0.476 0.495 -0.019[0.045] [0.035]
Type =1 (=1) 0.111 0.415 -0.304***[0.028] [0.034]
Type =2 (=1) 0.341 0.159 0.182***[0.042] [0.026]
Type =3 (=1) 0.040 0.179 -0.139***[0.017] [0.027]
Type =4 (=1) 0.508 0.246 0.262***[0.045] [0.030]
Head's age (years) 43.250 44.595 -1.345[1.034] [0.950]
Head's education (years) 2.468 3.304 -0.836*[0.325] [0.284]
Head is male (=1) 0.806 0.912 -0.106***[0.036] [0.020]
HH size 9.524 8.654 0.871*[0.430] [0.306]
Plot size (ha) 0.360 0.618 -0.258***[0.026] [0.036]
Owner (=1) 0.847 0.712 0.135***[0.032] [0.032]
Leaseholder (=1) 0.089 0.229 -0.141***[0.026] [0.029]
Sharecropper (=1) 0.016 0.024 -0.008[0.011] [0.011]
Others (=1) 0.040 0.034 0.006[0.018] [0.013]
Log asset value 4.213 4.835 -0.622***[0.099] [0.095]
F-test of joint significance 9.566***Number of Observations 124 204
37
Table 4. Changes in Outcome Variables by Treatment Status: Baseline, Mid-line, and End-line
Standard deviations are in brackets for mean values, and standard errors are in brackets for unconditional difference-in-difference. ***<0.01; **<0.05; *<0.1
Year 1
Treated ControlMean
differenceTreated Control
Meandifference
Treated ControlMean
differenceYear 2- Year 1 Year 3- Year 2
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)Panel AFertilizer (kg/ha) 214.071 254.340 -40.269 248.822 261.288 -12.466 232.750 255.110 -22.360 27.803 -9.894
[19.979] [32.869] [15.937] [17.609] [21.745] [17.994] [46.061] [36.572]Seed selection (=1) 0.906 0.864 0.042 0.929 0.978 -0.050 0.976 0.976 -0.000 -0.092* 0.050
[0.029] [0.034] [0.026] [0.015] [0.017] [0.017] [0.055] [0.040]Levelling (=1) 0.772 0.791 -0.019 0.857 0.677 0.180*** 0.867 0.810 0.058 0.199** -0.122
[0.040] [0.039] [0.036] [0.049] [0.037] [0.043] [0.081] [0.083]Canal/drainage construction/reparing (=1) 0.906 0.879 0.027 0.867 0.731 0.136** 0.855 0.929 -0.073 0.109 -0.209***
[0.028] [0.032] [0.034] [0.046] [0.039] [0.028] [0.071] [0.076]Transplanting in row (=1) 0.054 0.019 0.035 0.378 0.108 0.270*** 0.349 0.179 0.171** 0.235*** -0.099
[0.021] [0.014] [0.049] [0.032] [0.053] [0.042] [0.063] [0.090]Panel BRice Yield (ton/ha) 3.440 3.940 -0.499** 4.052 3.671 0.382 3.416 3.724 -0.307 0.881** -0.689
[0.164] [0.174] [0.238] [0.192] [0.203] [0.202] [0.387] [0.424]Gross output value (000 CFAF/ha) 603.159 669.393 -66.233 645.433 582.479 62.954 536.090 597.471 -61.380 129.187* -124.334*
[32.452] [31.737] [37.660] [31.545] [31.675] [31.760] [66.960] [67.446]Rice income (000 CFAF/ha) 405.308 405.091 0.217 413.726 353.458 60.268 232.174 292.198 -60.024 60.051 -120.292*
[31.311] [32.544] [36.667] [28.192] [38.128] [32.294] [64.909] [68.277]Rice profits (000 CFAF/ha) 331.539 320.196 11.343 243.209 230.344 12.864 108.260 155.150 -46.890 1.522 -59.754
[29.863] [34.177] [51.905] [32.074] [40.102] [31.341] [76.013] [81.382]F-test of joint significance 1.516 4.615*** 1.777*Number of Observations 101 103 98 93 83 84 395 358
Unconditional DIDYear 2 Year 3
38
Table 5. Estimated Results on the Dynamic Impacts of Training: Rice Productivity
Sample size is 353. Clustered standard errors at the production site level in parentheses. *** p<0.01, ** p<0.05, * p<0.1 Control variables included but not reported here: Year dummy, household size, head’s characteristics (age and its square, years of education and the male dummy), plot characteristics (cultivation size, tenure dummies for owner, leaseholder, sharecroppers), log household asset value, and local fixed effect. Attendance rate is instrumented by the treatment dummy, while attendance × end-line is instrumented by treatment × end-line dummy.
Rice Yield(ton/ha)
Gross outputvalue
(000 CFAF/ha)
Rice income(000 CFAF/ha)
Rice profits(000 CFAF/ha)
(1) (2) (3) (4)ITTTreatment (=1) 0.748* 140.105** 102.768* -8.362
(0.335) (51.348) (45.640) (50.222) × endline -0.642* -126.704** -52.889 69.781
(0.279) (44.488) (61.507) (70.156)Wald test (Ho: total effect is zero) 0.56 0.26 0.89 2.18R-squared 0.465 0.418 0.678 0.709TOTAttendance rate (instrumented) 1.453** 270.438*** 203.150** -15.249
(0.629) (94.168) (81.749) (85.727) × endline (instrumented) -1.254** -245.705*** -114.245 123.358
(0.527) (80.283) (96.329) (116.879)Wald test (Ho: total effect is zero) 0.78 0.37 1.13 2.71*R-squared 0.462 0.416 0.681 0.710
39
Table 6. Estimated Results on the Dynamic Impacts of Training: Agronomic Practice
Sample size is 353. Clustered standard errors at the production site level in parentheses. *** p<0.01, ** p<0.05, * p<0.1 Control variables included but not reported here: Year dummy, household size, head’s characteristics (age and its square, years of education and the male dummy), plot characteristics (cultivation size, tenure dummies for owner, leaseholder, sharecroppers), log household asset value, and local fixed effect. Attendance rate is instrumented by the treatment dummy, while attendance end-line is instrumented by treatment × end-line dummy.
Fertilzer(kg/ha)
SeedSelection
LevellingCanal
/drainageStraight-row
planting(1) (2) (3) (4) (5)
ITTTreatment (=1) 24.736 -0.033 0.178*** 0.119* 0.218**
(17.358) (0.031) (0.050) (0.062) (0.067) × endline -27.566 0.041** -0.202 -0.236** -0.227*
(24.593) (0.017) (0.134) (0.089) (0.099)Wald test (Ho: total effect is zero) 0.03 0.07 0.06 2.28 0.02R-squared 0.416 0.133 0.313 0.427 0.627TOTAttendance rate (instrumented) 47.437 -0.063 0.340*** 0.227* 0.417***
(30.788) (0.056) (0.084) (0.118) (0.095) × endline (instrumented) -52.111 0.076*** -0.380* -0.432*** -0.430***
(41.586) (0.030) (0.215) (0.154) (0.155)Wald test (Ho: total effect is zero) 0.04 0.08 0.06 2.81* 0.02R-squared 0.415 0.129 0.295 0.418 0.628
40
Table 7. Estimated Results on Heterogeneous Impacts of Training by Production Site Type
Sample size is 353. Clustered standard errors at the production site level in parentheses. *** p<0.01, ** p<0.05, * p<0.1 Control variables included but not reported here: Year dummy, household size, head’s characteristics (age and its square, years of education and the male dummy), plot characteristics (cultivation size, tenure dummies for owner, leaseholder, sharecroppers), log household asset value, and local fixed effect.
Rice Yield(ton/ha)
Gross outputvalue(000
CFAF/ha)
Rice income(000
CFAF/ha)
Rice profits(000
CFAF/ha)
Seedselection
Fertilzer(kg/ha)
LevellingCanal
/drainage
Straight-row
planting
(1) (2) (3) (4) (5) (6) (7) (8) (9)ITTTreatment 0.554* 86.740 75.152 35.290 -0.005 43.793*** 0.162*** 0.036 0.140
(0.245) (60.225) (73.338) (39.625) (0.018) (7.013) (0.026) (0.035) (0.076) × Type 2 0.379 47.181 47.018 -20.633 0.024 -36.936* -0.205* -0.173*** -0.151
(0.871) (115.163) (145.227) (103.333) (0.021) (17.298) (0.107) (0.047) (0.133) × Type 3 -0.185 11.013 7.947 75.635 0.082 -51.015*** -0.075 0.001 -0.014
(0.228) (73.335) (74.681) (69.812) (0.067) (11.668) (0.043) (0.108) (0.068) × Type 4 -0.840** -104.229 -52.751 -79.356* -0.131*** -67.815*** -0.103** 0.024 0.019
(0.263) (62.366) (71.359) (36.842) (0.019) (6.568) (0.035) (0.035) (0.070)R-squared 0.468 0.415 0.678 0.709 0.167 0.419 0.307 0.414 0.619
41
Table 8. Estimated Results on the Dyadic Regression
Clustered standard errors at the study site level in parentheses: *** p<0.01, ** p<0.05, * p<0.1
Control variables included but not reported here: The sum and differences of household size,
heads’ age, head’s years of education, cultivation land size, asset values, and a dummy equal to
one if the household heads are same gender as well as local fixed effects.
Ask agriculturaladvice =1
Both treat [Treat, Treat] 0.045*
(0.020)
× endline -0.018
(0.031)
Own treat, pair control [Treat, Control] 0.019
(0.030)
× endline 0.036
(0.040)
Own control, pair treat [Control, Treat] -0.063**
(0.027)
× endline 0.091**
(0.033)
N 2096
R-squared 0.063
42
Table 9. Estimated Results on the Linear-in-mean Model
Sample size is 144. Clustered standard errors at the production site level in parentheses. *** p<0.01, ** p<0.05, * p<0.1 Control variables included but not reported here baseline respondent’s and average values in respondent’s network of: household size, head’s characteristics (age and its square, years of education and the male dummy), plot characteristics (cultivation size, tenure dummies for owner, leaseholder, sharecroppers), log household asset value, and local fixed effect, as well as the treatment dummy for the respondent.
Rice Yield(ton/ha)
Gross outputvalue(000
CFAF/ha)
LevellingCanal
/drainage
Straight-row
planting
(1) (2) (4) (5) (3)The average outcome value (lagged) in network 0.279 0.206 0.273* -0.085 0.079
(0.197) (0.155) (0.133) (0.116) (0.070)
Network size -0.107 -12.155 -0.024 0.027* 0.025
(0.129) (20.075) (0.017) (0.012) (0.021)
Share of treatment in network 2.426 494.419* 0.537** 0.256 0.586**
(1.312) (207.006) (0.202) (0.224) (0.217)
R-squared 0.448 0.386 0.388 0.503 0.737
43
Appendix 1. Estimation Results for Non-attrition Probit Model
Sample size is 328. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1
Year 2 Year 3
(1) (2)
Head's age (years) -0.015 0.024
(0.043) (0.035)
Head's age squared (years) 0.000 -0.000
(0.000) (0.000)
Head's education (years) 0.037 0.021
(0.023) (0.022)
Head is male (=1) 0.570** -0.177
(0.283) (0.271)
HH size -0.011 -0.011
(0.021) (0.021)
Plot size (ha) 0.116 0.641**
(0.217) (0.270)
Owner (=1) -0.078 0.191
(0.376) (0.381)
Leaseholder (=1) -0.283 0.314
(0.412) (0.410)
Log asset value 0.194*** 0.081
(0.068) (0.067)
Constant -0.633 -2.175**
(1.142) (1.042)
Production site fixed effects Yes Yes
N 328 328
Top Related